2D Map and Scene Generator

Generate 2D Map

Build the smallest scene bundle the request needs, in the game's art, camera, renderer and data conventions. A background picture needs no game system; a playable map needs explicit geometry and data beyond a painting.

Use when

• A playable map, level, room set, tilemap, terrain or prop kit, with collision and exits.
• A parallax background, a battle, title or story backdrop, or an HD-2D plate with lights and motion.
• Map data must be validated, previewed or exported to Tiled, Godot or LDtk.

Do not use when

• Actors, items and attack FX: generate2dsprite. Character motion from video: video2dsprite.
• Code-drawn tiles, layouts or parallax the user asked for: codeart2d. Only an API or CLI call: generate2dmedia.

Capability check (once per session)

Run python "<skill-dir>/../generate2dmedia/scripts/forge_doctor.py" --host-tools <tools> --save <output>/doctor.json, where <tools> lists the media tools in your own tool list (image_gen, image_edit, image_to_video) or none. Its ROUTES block shows which API keys are configured (yes or no), the local readiness and the resolved order. If encoding.stdout fails, set PYTHONUTF8=1.

Art source

Image generation is the default for map art: terrain and tile sheets, prop kits, HD-2D plates, painted backgrounds and parallax layers. Map data (layout, collision, exits, spawns) is data you write and validate, not art. Record each asset's art_source (api, host_image, existing, or code only on request) and name the route in your reply. A route the user names explicitly wins.

1. API, when a key is configured: OpenAI (OPENAI_API_KEY), Google Gemini (GEMINI_API_KEY), xAI (XAI_API_KEY), BytePlus Seedream (ARK_API_KEY) or fal.ai (FAL_KEY), in the environment or the user config file (route-media.md). The configured key is the owner's consent, so there is no per-call question.
2. Local: your own image tool (Codex image_gen, Grok's native tool) when you have one, otherwise the user's signed-in Codex CLI, then Grok CLI.
3. codeart2d (autotiles, layouts, stylized parallax, ambient loops) only when the user explicitly asks for code-drawn art, or when route_media.py prints no-route (exit 3). Then say "code-drawn, no image model".

Every image goes through one command: python "<skill-dir>/../generate2dmedia/scripts/route_media.py" image --prompt-file <txt> --reference <png> --size 1536x1024 --out-dir <new>; masked plate motion uses route_media.py video. Each prints one JSON line with route, artifact, sha256 and estimateUsd. Existing art the user supplies is always welcome. Never claim a generation that did not happen; keep the prompt and contract and report a missing capability.

Host notes

• Codex: image_gen is your own image tool; with no configured key use it and look at results with view_image. An image this session made that is not in the project: python "<skill-dir>/../generate2dmedia/scripts/cli_media.py" adopt --codex-thread <thread-id> --output-dir <new>. codeart2d and generate2dmedia are explicit-only: open their SKILL.md when this one routes there.
• Claude Code: no built-in image generator: route_media.py takes the API or the local CLIs. Look at every PNG you make with Read (debug overlays, previews); run tools with Bash; <skill-dir> is ${CLAUDE_SKILL_DIR}.
• Grok: its native image and video tools are your own tools; use their real schema.

Commands

Run each tool as one line from the user's project root: python "<skill-dir>/scripts/<tool>.py" .... <skill-dir> is this skill's folder; sibling skills sit beside it (<skill-dir>/../codeart2d). Keep inputs and outputs inside the project. Every tool writes a new --output-dir (or a new --output file): it refuses an existing one, stages beside it and publishes only after its checks (--strict-qc / --strict publish nothing on failure). Success prints one JSON line; errors print error: ... and exit 1; usage errors exit 2. Needs Python 3.10+, numpy and Pillow (scipy recommended); ffmpeg 5.1+ for scene motion; node for the browser runtime. --help lists every flag.

Plan before art

Record viewport and logical resolution, world extent, origin, camera and zoom, actor scale and footprint, routes, landmarks, interactions, exits and render order. Source pixels are not world units: measure returned sizes and declare uniform scale and anchors. Choose by behaviour (map-strategies.md; starting values per genre in map-presets.md):

• scene_mode: authored area, town or arena; terrain plus props and explicit geometry.
• tile_mode / grid_mode: repeated or editable terrain or cell rules; tiles plus layer data.
• room_chunk_mode: connected rooms from a shared kit with door sockets.
• side_scroll_mode: side-view scenery or a playable platform stage with movement geometry.
• baked_scene_mode: a complete battle, title or story picture, with optional hotspots or bounded motion.

Split objects only when interaction, depth, reuse, editing or motion needs it. Pixel art needs deliberate logical pixels and nearest scaling; HD-2D pairs pixel actors with painted plates and light.